Model compression via pruning and knowledge distillation for person re-identification

Abstract

Person re-identification (ReID) is an important problem in intelligent monitoring. Recently, with the development of deep learning, convolutional neural networks have achieved state-of-the-art performance on person ReID problems. However, the deep neural network models used by these methods tend to have large number of parameters and high computational cost, thereby hindering their deployment on resource-constraint devices or real-time applications. In this study, we propose a method that distills the knowledge to a pruned model to reduce the parameters, which can be divided into two stages: one is to apply unstructured pruning method on over-parameterized models, whereas the other is to carry out representation and metric learning-based knowledge distillation on the model after pruning to improve performance. Finally, the proposed method can effectively reduce the total number of parameters by 8.4 with only 0.1% drop of rank-1 accuracy on the Market1501 dataset and no drop of rank-1 accuracy on the DukeMTMC-reID dataset.